Perretti et al. (2013a,b) showed that such a nonparametric
approach to forecasting could outperform parametric forecasts in
some stochastic dynamical systems in chaotic regimes, even
in situations in which the parametric forecast was based on the same
model that was used to generate the data. This poor performance of
parametric forecasting was due to two main reasons. The first reason
is the well known instability of parametric inference in chaotic
systems (Berliner, 1991; Wood, 2010). Hartig and Dormann (2013)
recalled an alternative method to calibrate the model which is based
on a segmentation of the time series to circumvent the sensibility to
initial conditions encountered in chaotic regimes (Pisarenko and
Sornette, 2004). Armed with this trick, Hartig and Dormann (2013)
have shown that the parametric forecast is more accurate than the
nonparametric one in the case where the same model was used to
generate the data and to ground the forecast